基于改进DenseNet网络的肺结节检测仿真

Simulation of Pulmonary Nodule Detection Based on Improved Densenet Network

针对目前基于神经网络的肺结节检测算法存在准确度低、耗时长的问题,提出了一种基于改进DenseNet网络的肺结节检测模型,通过在稠密神经网络的稠密块中采用分组卷积的方式来优化网络结构,减少网络参数的同时丰富了提取图像特征数量,避免卷积过程中梯度消失问题,提高了肺结节检测效率。之后将模型在LIDC-IDRI数据集上进行仿真,从参数量,准确率,AUC值三个方面进行评判仿真。仿真结果表明,与对比的基于深度神经网络的肺结节良检测模型相比,提出的方法提高了神经网络性能,有更高的分类准确率,准确率达到了92.4%。

Aiming at the problems of low accuracy and long time consuming in the current lung nodule detection algorithm based on neural network, a lung nodule detection model based on the improved Dense Net network is proposed, which optimizes the network structure by means of grouping convolution in the dense blocks of the dense neural network, reduces the network parameters and enriches the number of extracted image features, avoids the gradient vanishing problem in the convolution process, and improves the efficiency of lung nodule detection. Then the model was simulated on the LIDC-IDRI data set, and the evaluation experiment was conducted from three aspects: number of parameters, accuracy and AUC value. The simulation results show that the proposed method improves the performance of the neural network and has a higher classification accuracy of 92. 4% compared with the comparison of the lung nodules detection model based on the deep neural network.